| _version_ | 1866901191871430656 |
|---|---|
| author | HIDEKI |
| author_facet | HIDEKI |
| contents | <p>Large language models increasingly consume external reasoning traces (tool logs, multi-agent outputs), yet must infer trace reliability to choose between independent computation and cue extraction. We introduce a controlled corruption paradigm that independently manipulates trace degradation and embedded answer cues. Across 28 experiments (71,168 trials) with GPT-4o and Claude, we observe reliability-gated cue extraction. In verifiable arithmetic, GPT-4o's wrong-cue following increases monotonically with corruption (0.5% to 8.5% at λ=0.8), while Claude remains near zero. Vulnerability is consumer-determined, amplified by cue salience (up to 21%) and naturalistic tool-log traces (29.1%), and does not require explicit "Answer:" markers. In unverifiable knowledge tasks, both models show elevated cue dependence (~14%). Verification-oriented prompts suppress cue following to 0%, providing a practical control principle for trace-consuming systems.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18254275 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Reliability Inference Drives Cue Extraction in Large Language Models Consuming External Reasoning Traces HIDEKI large language models chain-of-thought reasoning reliability inference prompt injection AI safety LLM evaluation multi-agent systems retrieval-augmented generation <p>Large language models increasingly consume external reasoning traces (tool logs, multi-agent outputs), yet must infer trace reliability to choose between independent computation and cue extraction. We introduce a controlled corruption paradigm that independently manipulates trace degradation and embedded answer cues. Across 28 experiments (71,168 trials) with GPT-4o and Claude, we observe reliability-gated cue extraction. In verifiable arithmetic, GPT-4o's wrong-cue following increases monotonically with corruption (0.5% to 8.5% at λ=0.8), while Claude remains near zero. Vulnerability is consumer-determined, amplified by cue salience (up to 21%) and naturalistic tool-log traces (29.1%), and does not require explicit "Answer:" markers. In unverifiable knowledge tasks, both models show elevated cue dependence (~14%). Verification-oriented prompts suppress cue following to 0%, providing a practical control principle for trace-consuming systems.</p> |
| title | Reliability Inference Drives Cue Extraction in Large Language Models Consuming External Reasoning Traces |
| topic | large language models chain-of-thought reasoning reliability inference prompt injection AI safety LLM evaluation multi-agent systems retrieval-augmented generation |
| url | https://doi.org/10.5281/zenodo.18254275 |